Muse Spark Open: Meta's 816 billion parameter open-weight model arrives without scores

Meta released Muse Spark Open on October 15, 2026, an 816 billion parameter open-weight model with 42 billion active parameters and a 1 million token context window. Meta calls it "the most capable US open-weight model to date," but it published no scores in the announcement, and its strongest model stays closed.

The release splits Meta's strategy in two. The frontier stays behind Meta's own doors, while a weaker sibling goes out under a restrictive license. Here is what the launch page does and does not establish.

What Meta actually released

Per Meta's research blog, Muse Spark Open is a natively multimodal mixture-of-experts model with 816 billion total parameters and 42 billion active. Meta says it is built for coding, agentic work and multi-step reasoning.

Weights are available through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. Hosted APIs come from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Supported inference engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth.

The headline demonstration is local. Meta says the model runs on two Apple M5 Ultra Mac Studios using MLX, each with up to 512 GB of unified memory, so the pair behaves like one machine with about a terabyte. Meta showed on-device coding in an open-source iOS app and video editing in DaVinci Resolve.

The license is not an open source license

Meta releases the model under the Muse Model License, which it describes as having limited commercial terms while letting many developers use the model free of charge. That is a custom license with conditions, not an OSI-approved one. The Open Source Initiative's definition requires training data information and code, not just weights.

The Next Web, reporting the launch, noted the earlier Glimmer model, a 30 billion parameter model distilled from Muse Spark, shipped in August under Apache 2.0. Muse Spark Open is a step down in permissiveness from that, even as it is a step up in size.

What is missing: scores

Meta's page links to an evaluation methodology report but lists no benchmark scores. It says a full evaluation report is "available soon." That means "most capable US open-weight model" is, for now, a claim without a published table behind it.

Meta also says it added extra safety post-training because of the risks of releasing weights, and that the model met its release thresholds under the Meta Superintelligence Scaling Framework, with details in the Hugging Face model card. Gating a release on a framework Meta wrote itself is a commitment, not an audit. Our look at Meta's Muse Wonder launch found the same pattern: a strong claim, no numbers.

The strategy: open at one tier, closed at the top

Alexandr Wang, Meta's chief AI officer, told reporters: "Because of both safety and capabilities, we will generally not open-source our most powerful models," as reported by The Next Web. He added that Meta intends to keep releasing good open models, and said he expects most of the world's AI to run on open models that should be American, with Chinese developers building on Muse Spark Open.

Meta says Muse Spark Open is less capable than Muse Wonder, the frontier model it launched on Monday, October 12. So Meta now runs a tiered model: the top model stays closed, and a weaker one goes out as weights.

The EU angle is live. The Next Web notes that Article 53 of the EU AI Act exempts genuinely free and open models from some documentation duties, but not models with systemic risk. Meta declined to sign the EU code of practice, and the Commission can fine up to 3 percent of global turnover. Whether this model counts as open under that test is a question regulators, not Meta, will answer.

Who gets something real from this

Three groups benefit first, and the list is short.

Developers with a big workstation get a large local model. Cloud customers get another option next to Mistral Large 4, which came out with a trillion parameters and open weights. And Meta gets distribution: if Chinese labs and Western startups build on its weights, the ecosystem leans on Meta's stack.

What nobody outside Meta can say yet is how it compares. A 42 billion active parameter model can be very good, but the only public test of that is independent evaluation, and none exists today.

How to judge it when scores arrive

When Meta's evaluation report lands, check four things. First, which models it is compared against, and whether they are the strongest open-weight models or older ones. Second, whether the tests are public benchmarks or internal ones. Third, whether anyone outside Meta ran them. Fourth, whether the numbers cover coding and agentic tasks, since those are the uses Meta pitched. A model built for agents should be tested on long, multi-step tasks, not only short questions. Until those four answers exist, the practical advice is simple: try it on your own workload using one of the hosted APIs before buying hardware.

Our take

This is a real release with a smaller claim than its marketing. Weights are out, the deployment story is unusually broad and the two-Mac-Studio demo is concrete. But no scores, a custom license and a closed frontier mean we would not call it a win for open AI yet. Wait for the evaluation report and for Artificial Analysis or Epoch to run it, then judge it against the open models already out.